feat(azure/rhel-ai): add GPU guardrails for instance selection

- Add isGPUCapableSize helper matching ND/NC series (NV excluded)
- Shallow-copy ComputeRequestArgs before mutation to avoid caller side-effects
- Default ComputeRequest.GPUs to 1 so filterCPUsAndMemory auto-selects
  only GPU-capable instance types when no explicit GPU count is set
- Warn when caller explicitly provides compute sizes that are not
  GPU-capable (expected ND/NC series; vllm requires a GPU device)

Author: rishupk

pkg/provider/azure/action/rhel-ai/rhelai.go

@@ -34,16 +34,45 @@ func imageId(accelerator, version string) string {
 	return imageIdFromName(fmt.Sprintf(imageNameRegex, accelerator, version))
 }
 
+// isGPUCapableSize returns true for ND-series and NC-series Azure VM sizes,
+// which are the compute GPU families supported for RHEL AI workloads.
+// NV-series (visualization GPUs) is intentionally excluded.
+func isGPUCapableSize(vmSize string) bool {
+	lower := strings.ToLower(vmSize)
+	return strings.HasPrefix(lower, "standard_nd") || strings.HasPrefix(lower, "standard_nc")
+}
+
 func Create(mCtxArgs *maptContext.ContextArgs, args *apiRHELAI.RHELAIArgs) (err error) {
-	logging.Debug("Creating RHEL Server")
+	logging.Debug("Creating RHEL AI Server")
 	sharedImageID := imageId(args.Accelerator, args.Version)
 	if args.CustomImage != "" {
 		sharedImageID = imageIdFromName(args.CustomImage)
 	}
+	// Shallow-copy to avoid mutating the caller's ComputeRequestArgs.
+	computeReq := *args.ComputeRequest
+	// Ensure GPU-capable instance selection for auto-selection paths.
+	if computeReq.GPUs == 0 {
+		logging.Debug("RHEL AI: GPUs not set, defaulting to 1 for GPU-capable instance selection")
+		computeReq.GPUs = 1
+	}
+	// Warn when the caller explicitly specifies compute sizes that are not GPU-capable.
+	if len(computeReq.ComputeSizes) > 0 {
+		allNonGPU := true
+		for _, s := range computeReq.ComputeSizes {
+			if isGPUCapableSize(s) {
+				allNonGPU = false
+				break
+			}
+		}
+		if allNonGPU {
+			logging.Warnf("RHEL AI: none of the specified compute sizes %v appear to be GPU-capable "+
+				"(expected ND-series or NC-series); vllm may not start without a GPU", computeReq.ComputeSizes)
+		}
+	}
 	azureLinuxRequest :=
 		&azureLinux.LinuxArgs{
 			Prefix:         args.Prefix,
-			ComputeRequest: args.ComputeRequest,
+			ComputeRequest: &computeReq,
 			Spot:           args.Spot,
 			ImageRef: &data.ImageReference{
 				SharedImageID: sharedImageID,